Query Optimization Detection Transformer for Small Objects in Remote Sensing Images

Object detection in remote sensing images is a challenging task. Remote sensing images contain substantial background noise and complex contextual information, which weakens feature representation of small objects, making detection difficult. To tackle these problems, a detection Transformer for small objects in remote sensing images is proposed, called QO-DETR. To strengthen feature representation of small objects, a query proposal generation module is designed to select queries based on multi-class classification scores. These queries provide the initial position embeddings for object queries in the decoder, enabling the decoder's attention mechanism to focus on object regions. To make the model more robust to noise, a group denoising module is developed to boost the reconstruction of object features from noise. To accurately locate small objects, a query cascade refinement strategy is designed, and each decoder layer refines anchor parameters under the guidance of preceding layers, and the spatial alignment of anchor boxes and objects is achieved layer by layer. Experiments have been carried out on DIOR. AP and APS on DIOR reach 51.3% and 13.4%. QO-DETR shows superior performance in detecting small objects.